View Synthesis for Multiview Video Compression

We consider multiview video compression: the problem of jointly compressing multiple views of a scene recorded by different cameras. To take advantage of the correlation between views, we compare the performance of disparity compensated view prediction and view synthesis prediction to independent coding of all views using H.264/AVC. The proposed view synthesis prediction technique works by first synthesizing a virtual version of each view using previously encoded views and using the virtual view as a reference for predictive coding. We present experimental coding results showing that view synthesis prediction has the potential to perform significantly better than both disparity compensated view prediction and independent coding of all views.